Decentralized AI is being repriced after the Fable 5 incident
Source:The Defiant
Compiled by Yuliya, PANews
Original title: Frontier AI's Game of Thrones and the Debate of Decentralization: Looking at the Future of DeAI from the Fable 5 Blocking Scandal
Editor's note: Last week, Anthropic's release of Claude Fable 5 unleashed the most severe crisis of trust in the cutting-edge AI field: researchers discovered that once the model suspected that users were developing a competitor, it would “secretly poison” the quality of responses. In addition, the model had a 30-day data retention requirement, causing it to be disabled within Microsoft. This raises the question the crypto sector has been asking for years: Should any single company control so much cutting-edge AI?
In response, Camila Russo, founder and CEO of The Defiant invited CoinFund founder Jake Brukhman, Sentora and The Sequence founder Jesus Rodriguez, and Dragonfly managing partner Haseeb Qureshi to have a heated debate on the future direction of decentralized AI.
Big Model Wars, Open Source Trends, and “Lockdown” Panic
Haseeb: Our current investment logic is: in the future, everyone will see more and more “non-cutting-edge” models emerge, and users' spending on model tokens (computing power expenses) will also increasingly flow to these non-cutting-edge fields. Everyone knows that spending money on big, cutting-edge models is unsustainable, and the vast majority of people simply can't use that level of intelligence.
There are many distilled, open source, or open weighted models on the market. The price is very affordable, and you can completely assign different tasks to them. There's a saying on the internet that people actually use Mythos or Claude Fable 5 level models to rename a file — this will happen more and more as we become more familiar with these models. What you need to think about is: Should I use a knife to kill a chicken?
Having said that, the term “decentralized AI” is too broad. If it only refers to “all kinds of models developed by different agencies” (such as the OpenRouter model), then this is no different from our current world. But if it refers to “using a decentralized network to train or run AI models,” then that's another set of logic. We are actually quite pessimistic about the latter. Currently, we don't see any reliable reason to prove that the economic benefits and market demand for training or operating models in a decentralized environment have been established.
Of course, the way Fable was released this time did cause a strong backlash. People have a sense of possessiveness about good products, and once they use them, they feel “don't want to take it away unless I die”. When the government suddenly stepped in to block it, everyone must have felt deprived. But at the same time, if you remember the scene when Mythos was first released, it was terrifying — in front of it, all of our existing software, operating systems, or browsers were just as vulnerable as Swiss cheese. No one popped up at the time and said, “You should open it to all.”
Some people say the US government is acting crazy here. Anthropic claims they have cleared up all the concerns of the NSA before releasing Fable 5, but as far as I understand it, the NSA has long since stepped in to block Mythos. Mythos was promoted to just over 30 partners in Project Glasswing, and these partners were carefully selected by the government rather than by Anthropic. So the statement “Fable was released without the government” is clearly untenable. Rumor has it that Amazon's president Andy Jassy went to the government or the White House and told them that the model had a jailbreak vulnerability before the government realized the danger and immediately blocked Fable 5 across the US.
This governance and security mechanism is clearly inadequate. While I agree that what is happening in the lab (whether it's Anthropic or OpenAI) is extremely dangerous and requires careful treatment, I also believe there is huge economic value in the distribution of open source and open weighting models, and the two must evolve in parallel.
*Note: Project Glasswing is a cybersecurity project initiated by Anthropic and promoted by a number of technology companies. It was launched in April 2026.
Jesus: Without talking about the apocalyptic nature of technology, I did hear from people in the cybersecurity industry that Mythos is really scary. I talked to some people at Anthropic after it was released, and the question was very real. But I've heard well-known CEOs in the cybersecurity field say that they would prefer to have open access to this model because releasing it directly would give all these security companies enough time to prepare. If you try to limit it or delay the release for three months, you'll never get enough room to buffer. But the opposite view is: if Mythos were released directly, would it have disastrous consequences?
Haseeb: We are in the blockchain field. If North Korea gets this model, do you really think it won't be disastrous?
Camila: But isn't there an argument: if everyone has it, can the risk be reduced because everyone can test?
Haseeb: Not everyone has nuclear weapons.
Jake: It's not appropriate to use a nuclear weapon as an analogy. In the case of Mythos, it's a model that explores system vulnerabilities. We need to make a financial calculation: hackers spend money on Mythos to find bugs, and website owners also spend money to defend against them. Is this market really equal? Do hackers really think it's a good idea to spend a lot of time working on a Linux system bug that can't be monetized at all?
If this model of exploiting vulnerabilities is in the hands of only a few people (for example, large companies can use it, ordinary people can't), you're actually creating an imbalance. Some are able to protect their assets, while others can only be beaten. So I personally think it would be better if everyone had equal access to the model.
I'm not a cyberpunk rebellious spirit; this is an inevitable trend in the market. Today you see cutting-edge closed source models, but there are also a large number of open source models (mostly created by Chinese laboratories). Although they are at a disadvantage in terms of computing power, the gap between them and cutting-edge models in terms of various evaluation indicators is only a few percentage points. Epoch.ai's chart clearly shows that the gap between open source and closed source models is rapidly narrowing. Even if Anthropic wants to be the “big brother” to protect everyone, the reality is that people need these models to protect their websites and software. They always get their hands on it—either from Anthropic, open source from an Asian lab, or trained on a decentralized network.
The boundaries of export control, regulation and free access
Camila: Jake, don't you think there should be any fences at all? Should it be completely open to everyone?
Haseeb: Let me add to that question. Do you think “export control” should not exist as a concept at all? Because apart from AI, the network itself is an element of war.
Jake: I have no position on politics. I'm just a technical person, and I don't work at the State Department. If the US government decides to implement export controls, that is their business. But it's two different things from “whether technology should be shared around the world.”
Assuming Fable was trained on a decentralized network, no one has full model weights (some in the US, part in Amsterdam, part in Australia). If the US imposes export controls on that portion of its domestic weight, this model may well be replicated elsewhere in the world. This is the problem with the US enforcement mechanism. Look at Bitcoin, it's a currency with independent sovereignty, decentralized, and no one can stop it. Haseeb just said he wasn't sure if the market needed decentralized AI. This is like saying “I don't know if anyone needs PoW (Proof of Work)” in 2011. In fact, because everyone is in demand for a global, unlicensed currency, the demand for technology is huge. Similarly, there is a huge demand for global, unlicensed AI, and the US State Department can't stop it, whether it likes it or not.
Jesus: Regarding the export control analogy, what if you restricted everyone's access to Mythos, but a model with open weights suddenly evolved cyber attack capabilities on its own? Looking at the current cybersecurity benchmarks, DeepSeek-v4 or Qwen 3.7 rank very high. It's only a matter of time before these models become capable of cyber attacks.
The AI community likes to use nuclear weapons as an example: four years after World War II, America possessed nuclear weapons while the rest of the world did not. There is a theory that if the US were pressured at the time, communism might never have developed in Eastern Europe. But later, the Soviet Union also developed nuclear weapons. What bothers me is not opening it up to everyone in the first place, but about selectively deciding who can visit. If this is an export control, why doesn't every US company have access to it?
Haseeb: Regarding Fable, we need to sort out the details. The government is calling for Fable to be shut down for all non-Americans. Currently, Anthropic does not have sufficient KYC (real-name authentication) mechanisms to guarantee that they can comply with this, and export control is a strict accountability system. If the model falls into the hands of non-Americans, you're in trouble. That's why they don't currently have the confidence to say they can do it. Currently, Polymarket predicts that the probability that they will be able to resume operations for Americans by the end of July is 77%, while the probability of recovering around the beginning of June is about 50%.
Obviously, the idea of “banning any foreigner from using Fable 5” is outrageous in itself. The US has a large number of foreign employees with H1B visas. If you have French engineers in your programming team, they aren't allowed to use Fable, which is ridiculous. This is likely to be changed by negotiations before actual implementation, and if Anthropic can fix the bugs and implement stricter controls, it may not be necessary to completely shut down non-US actors.
But that's not the case with Mythos. FFable was originally supposed to be just a “good model” for writing code and drafting emails, but in the face of Mythos, the US government's attitude was: No, this stuff can only be given to Americans, and “only to the people we name on the list.” This isn't an export control anymore; it's simply an AI version of the “Manhattan Project.”
As far as I know from reliable sources, the government led the Project Glasswing process, which is why it was all big companies like Microsoft that got places, not some random cybersecurity company. This is not surprising to a government that sees it as an extremely dangerous offensive cyber weapon; we also handle fighter jets and missiles in the same way. This isn't just what Anthropic wants 30 companies to use for commercial marketing purposes; they can't help but use their products all over the world.
Camila: In the field of cryptography, we're seeing a sharp increase in the number of AI-generated hacking attacks, and we can infer how risky Mythos would be if it were widely adopted. Jake, do you think it's reasonable to restrict certain groups from using these models in certain circumstances? Or are you still insisting they should be open to everyone?
Jake: As I said, this and “whether decentralized AI technology works” are two separate issues. The government can of course enact laws to regulate it; this is not a black and white choice. However, decentralized technology can bring more competition by lowering the entry threshold. It uses commodity grade hardware to reduce costs.
I'm talking to a founder today who is reasoning on heterogeneous commodity GPUs, and he thinks this will be a cheaper option for consumers in the long run as electricity costs rise. At the end of the day, all technological advancements are aimed at reducing costs and barriers. AI can be said to be the most centralized industry in the world at present, and it needs to break the threshold the most. We support decentralized AI to protect consumers' right to choose, and ultimately to defend democracy.
Physical bottlenecks and algorithm breakthroughs in decentralized AI
Camila: What would happen if only a few centralized companies eventually controlled most of the AI models used in the world? If there isn't actually a successful decentralized AI, what's the cost?
Jesus: I have to refute Jake. From a technical point of view, using a decentralized approach to create a Mythos-level model is definitely much more expensive than centralization. Nvidia has a deep-water moat that few people mention: with the exception of Google, which has TPU, all large-scale architectures currently run on hundreds or thousands of Nvidia GPUs, and AMD doesn't have this actual combat data at all.
I actually support centralized AI; I have set up two companies in this field. Decentralized AI is nothing new; it has never found a product market fit (PMF) before. Previously, because the model was small and simple enough, decentralization didn't make much sense. Now that they're big enough, decentralization has become very difficult. Moreover, we all have gaps in talent, pay, and funding. Much of the inference is actually not done on the most advanced GPUs; on previous generation GPUs, pre-training only required H100.
Jake: The supply of GPUs has been at a bottleneck for the past few years, and prices have continued to rise. In 2026, it will be very difficult for the average mid-market startup to find the H100. Large-scale training in history has been conducted in luxury data centers that require nuclear power support — those industrial-grade GPUs have 132 GB of memory, and the bandwidth between nodes reaches 1 to 3 Tb/s. If I told you we could transfer this process to a consumer device (like a regular Nvidia GPU, or even your Macbook or Mac Studio) and run it on a normal consumer network, you'd say I'm crazy.
However, when faced with such huge computational demands, people are extremely motivated to change training methods and optimize algorithms. Let me use a quantum analogy: Google has two types of quantum experts. Those working with hardware say quantum computers can't solve any problems in ten years, while those working with software say “you have to be careful with Ethereum within 3 to 5 years.” Haseeb and Jesus are looking at the problem from a hardware perspective, and I'm looking at it from the perspective of someone who optimizes how the hardware is used.
We are making tremendous progress. Not only is research showing that post-reinforcement learning training can be 10x faster and cheaper, the Pluralis currently being run solely on the RTX 4090, which will show that you can train a real big language model (LLM) on pure consumer-grade devices. Because half of the data center's TCO (total cost of ownership) is facility maintenance and cooling, and device clusters (Swarm) don't have these costs, it will become cheaper. Even if it's slightly slower, it's still economically viable due to its much lower cost.
The earliest algorithms (such as DiLoco, Sparse LoCo, and Google's algorithm two years ago) increased the parameter scale from 10 billion and 40 billion to 72 billion. Macrocosmos has now reached the 100 billion parameter. The next generation of algorithms will break the model, and I think we'll reach trillion-level parameters using these methods.
Haseeb: Let me play the skeptic.
First, there are two limitations to building large models: computation and bandwidth. The laws of physics cannot be broken. If you don't physically place the devices together and communicate via a high bandwidth interconnect, but instead communicate through the public Internet and compress gradient updates, you will have to pay a huge price. Moreover, it is impossible to accurately calculate the “total cost of ownership (TCO)” of machines in a decentralized network, one in the east and one in the west. People involved in decentralized storage back then also said, “Although it's slow now, it would be nice if the algorithm was optimized later.” What about the results? There's no need for decentralized storage because it's neither cheap nor efficient when calculated.
The most important point is that the biggest limitation for training a large model is data. To train a model with a rough estimate of 8 trillion parameters like Mythos or Fable, you need massive amounts of token data. OpenAI and Anthropic spend huge sums of money to generate data from vendors. They are expensive to generate synthetic data and obtain user data from usage traces of Claude Code and Codex. Decentralized people don't have this data at all.
Leave the economics aside and look at the demand side. I think the core value proposition of cryptocurrencies is not decentralization; decentralization is just a means to be self-sovereign and resistant to censorship. This is why Satoshi Nakamoto designed Bitcoin. In the field of AI, what do people care about? The first is cost; the second is that you own the model and the data is not included in the training set; and the third is resistance to censorship. People hate Fable 5's censorship system and its internal mechanisms that secretly reduce performance.
Check out Venice AI, which is currently the darling of the crypto AI product industry. It uses Near AI for confidential calculations, protects privacy, and has zero data retention. However, the most commonly used model on Venice is not a model with decentralized training (not from Pluralis, etc.), but an open source weight model run by regular companies like DeepSeek or GLM-5. This shows that the development direction of AI may be: people want a private and censorship-resistant experience, but it doesn't have to be achieved through an underlying decentralized mechanism like Bitcoin or Ethereum.
Jesus: People with decentralized and centralized AI are often solving the problem of being two generations behind. A researcher told me the other day, “Pre-training hasn't been completely solved yet, but it's very boring.” A lot of innovation in reasoning comes from post-training, and now we're talking about recursion, continuous learning, etc. The gap between centralized AI is actually getting wider and wider due to the downsizing of talent and capital. As for small models and end side calculations, it is often very easy to use the big model directly as a distillation (such as Google's Gemma). If you set up a decentralized cluster and practice hard for a month, if any computer in the middle goes down and causes a complete crash, I don't even know how you'll end up.
Jake: You're wrong about that. Decentralized training clusters are extremely resilient to pressure. In giant data centers, if a GPU breaks down, you may need to restart training; in Swarm, GPUs of different sizes and shapes can enter and leave the network at any time during training without negative effects. The biggest proof is that Google recently stated on their blog that they are starting to use DiLoco-style algorithms in their data centers.
As for the data issue, Haseeb was right, but that doesn't mean that decentralized people don't have data; centralized people have data. There are many customers in the market who want better AI economics. For example, Kirkland & Ellis (Kirkland & Ellis) recently announced that they will spend $500 million to buy their own proprietary data sets for training, and they are even hiring AI engineers within the firm. For customers like them who have a $500 million budget and want to train their own models, decentralized networks eliminate data center cooling and maintenance costs, and as a computing floor, the costs will be greatly reduced.
Haseeb: The reason KY & E is doing this is because they don't want to share their data. If they put their data in a decentralized network, their data is exposed. They didn't do this because they felt they were good at training models; they wanted to internalize value. Why hand it over to Harvey (AI legal tool)?
Jake: Who said decentralized training must be open and transparent? It's totally possible to set up a privacy license system. More importantly, when the model's weight is distributed and no single entity holds all the weight, the model's users must pay for the network. This revenue stream no longer goes to OpenAI's Sam Altman or Anthropic's Dario, but to the network's token holders, buyers, and training participants. This gave the model a business model and revenue stream. Traditional law firms may not adopt it right away, but there are definitely companies that will find it a good way to finance products.
Cyber attacks, geopolitics, and the last bastion
Camila: If all of this can be achieved, decentralized AI is as powerful as centralized AI. Can decentralized networks resist censorship in a situation like the Fable model being shut down at the request of the government?
Jake: Resist censorship isn't actually a priority for these networks. But if you really want to do this, you can break up the neural network and spread the weight across dozens of countries, then you won't be able to be forced to do it. But I repeat that the ultimate goal of decentralized AI is to lower the threshold, democratize computing power, and enable more people to train models.
Jesus: OpenAI previously mentioned that “the model itself is no longer a product.” In the field of decentralized AI, people seem obsessed with building models, and are actually two or three generations behind existing technology. We should look for value in the infrastructure surrounding the model: environmental capabilities such as sandboxes for code execution and computation, evaluation mechanisms (Evals), synthetic data pipelines, etc. Many modern financial applications can be built on the intersection of DeFi and AI, but we haven't taken full advantage of them.
Haseeb: Back to the original question, what would happen if cutting-edge AI were to be completely open sourced and run all over the place, and even export controls couldn't stop it?
I think a “COVID-level” cybersecurity tsunami will break out around the world. Software that can't be patched and servers of small companies will be bombarded and left behind. Just look at the data on the chain: April 2026 was the month with the highest number of hacking attacks in crypto history, followed by May, which broke a new record. Although the total amount of money stolen is not an exaggeration, the frequency of incidents has skyrocketed, meaning that storing money in small agreements is more dangerous than ever before.
If everyone in the world were to carry a “bazooka,” a large amount of infrastructure would inevitably be destroyed. Although our system will be equipped with “tank armor” within two to three years after experiencing pain, the painful period during this period will be extremely severe.
Camila: Wouldn't it be better to have such a powerful tool in everyone's hands than just a few companies and governments?
Haseeb: In the “everyone” you mentioned, North Korea was included. Do you really want North Korea to take Mythos?
Camila: So would you rather leave the US government alone, or even let them censor others, rather than share it with all of us?
Haseeb: If I could only choose between “only the US can use it” and “the whole world can use it,” I would choose the US. If you really believe that AGI (General Artificial Intelligence) will arrive, then it is the most powerful weapon in human history. Since ancient times, weapons of mass destruction have all been controlled by sovereign countries; this is normal. I'm not worried about the Chinese government getting Mythos; they are acting steady and have long-term plans; I'm worried about North Korea, terrorists, and rogue hacker groups. Just like I'm not worried about China having a nuclear bomb, but I'm really worried about North Korea pressing a button.
Camila: Finally, Jake and Jesus can summarize. Haseeb's firepower is so strong that we need a little bit of decentralized faith to recharge.
Jake: From an investor's point of view, it's about finding areas with excellent risk-reward ratios. Decentralized AI is a really cool field. At dinner the other day, a friend of ours said, “Cryptocurrency is becoming a simple traffic business, what should we do?” In this world, decentralized AI can be said to be the last bastion in the cryptocurrency field, and it is cutting-edge technology that really works. I'm very excited about the companies we work in this field (such as Pluralis, Prime, Intel, Jensen, Bagel, Pearl, etc.).
Jesus: Decentralized AI is definitely valuable, but I'm still not optimistic about decentralized “pre-training.” I think there are huge opportunities in decentralized AI infrastructure. Crypto's underlying technology stack is too old, and the world is using AI to modernize and upgrade. The combination of DeFi and AI is definitely the next frontier.
Twitter:https://twitter.com/BitpushNewsCN
Compare the TG exchange group:https://t.me/BitPushCommunity
Compare TG subscriptions:https://t.me/bitpush
